Masked diffusion models have emerged as a powerful framework for text and multimodal generation. However, their sampling procedure updates multiple tokens simultaneously and treats generated tokens as immutable, which may lead to error accumulation when early mistakes cannot be revised. In this work, we revisit existing self-correction methods and identify limitations stemming from additional training requirements or reliance on misaligned likelihood estimates. We propose a training-free self-correction framework that exploits the inductive biases of pre-trained masked diffusion models. Without modifying model parameters or introducing auxiliary evaluators, our method significantly improves generation quality on text-to-image generation and multimodal understanding tasks with reduced sampling steps. Moreover, the proposed framework generalizes across different masked diffusion architectures, highlighting its robustness and practical applicability. Code can be found in https://github.com/huge123/FreeCorrection.

Diffusion Language Models (DLMs) have recently achieved substantial progress in natural language generation tasks. Recent research demonstrates that adaptive token generation ordering can significantly improve performance in mathematical reasoning and code synthesis applications. In this work, we investigate the optimization of generation order for both text-to-image synthesis and multimodal understanding. We first establish that, unlike structured problems in language generation such as Sudoku puzzles, model logits alone are insufficient for determining optimal generation sequences in text-to-image generation and multimodal understanding. To address this challenge, we introduce a learnable control module trained via Group Relative Policy Optimization (GRPO) to determine the generation order. Our results demonstrate that learning this control block substantially improves both text-to-image alignment and multimodal understanding in DLMs. In particular, it enhances the model's ability to capture fine-grained spatial relationships in generated images while also strengthening performance on multimodal reasoning and comprehension tasks. We evaluate our framework on GenEval, an object-focused benchmark for text-to-image alignment, where it achieves 4.08

Vision Language Models (VLMs) demonstrate strong perceptual abilities but remain limited in tasks requiring analytical reasoning across multiple visual states, such as multi-image comparison, change detection, and multi-step visual inference. These capabilities are critical for real-world multimodal applications where reasoning must be grounded in systematic differences between visual contexts. However, existing benchmarks rarely require both explicit visual comparison and analytical reasoning, leaving this capability underexplored. To address this gap, we introduce SD-MAR (Synthetic Data for Multi-image Analytical Reasoning), a framework for training and evaluating VLMs on multi-image analytical reasoning. SD-MAR constructs paired visual scenarios through controlled perturbations and generates reasoning tasks spanning semantic change attribution and quantitative comparison. We further train VLMs using GRPO-lite with Backward Discounted Allocation (BDA), a reinforcement learning approach that removes KL regularization to encourage stronger policy optimization while allocating greater credit to the later reasoning steps where analytical conclusions are formed. Experiments on Qwen2.5-VL-7B and InternVL3-8B show that GRPO-lite fine-tuning on SD-MAR improves in-domain accuracy by up to 36.95

Toxicity text detectors can be vulnerable to adversarial examples - small perturbations to input text that fool the systems into wrong detection. Existing attack algorithms are time-consuming and often produce invalid or ambiguous adversarial examples, making them less useful for evaluating or improving real-world toxicity content moderators. This paper proposes an annotation pipeline for quality control of generated toxic adversarial examples (TAE). We design model-based automated annotation and human-based quality verification to assess the quality requirements of TAE. Successful TAE should fool a target toxicity model into making benign predictions, be grammatically reasonable, appear natural like human-generated text, and exhibit semantic toxicity. When applying these requirements to more than 20 state-of-the-art (SOTA) TAE attack recipes, we find many invalid samples from a total of 940k raw TAE attack generations. We then utilize the proposed pipeline to filter and curate a high-quality TAE dataset we call TaeBench (of size 264k). Empirically, we demonstrate that TaeBench can effectively transfer-attack SOTA toxicity content moderation models and services. Our experiments also show that TaeBench with adversarial training achieve significant improvements of the robustness of two toxicity detectors.

Jailbreaking large-language models (LLMs) involves testing their robustness against adversarial prompts and evaluating their ability to withstand prompt attacks that could elicit unauthorized or malicious responses. In this paper, we present TurboFuzzLLM, a mutation-based fuzzing technique for efficiently finding a collection of effective jailbreaking templates that, when combined with harmful questions, can lead a target LLM to produce harmful responses through black-box access via user prompts. We describe the limitations of directly applying existing template-based attacking techniques in practice, and present functional and efficiency-focused upgrades we added to mutation-based fuzzing to generate effective jailbreaking templates automatically. TurboFuzzLLM achieves ≥ 95% attack success rates (ASR) on public datasets for leading LLMs (including GPT-4o & GPT-4 Turbo), shows impressive generalizability to unseen harmful questions, and helps in improving model defenses to prompt attacks.

Large Language Models (LLMs) deployed as autonomous agents with tool access present unique safety challenges that extend beyond standalone model vulnerabilities. Existing red-teaming frameworks like AgentHarm use static prompts and hardcoded toolsets, limiting their applicability to custom production systems.We introduce a dual-component automated red-teaming framework: AgentHarm-Gen generates adversarial tasks and evaluation functions tailored to arbitrary toolsets, while Red-Agent-Reflect employs iterative prompt refinement with self-reflection to develop progressively more effective attacks.Evaluating across 115 harmful tasks (71 generated, 44 from AgentHarm) spanning 8 risk categories, our method achieves substantial improvements: up to 162% increase in attack success rate on o4-mini and 86% success on Gemini 2.5 Pro. Successful attacks systematically decompose adversarial objectives into benign-appearing sub-tasks that circumvent safety alignment, highlighting the need for agent-specific guardrails.

As large language models (LLMs) become increasingly prevalent, ensuring their robustness against adversarial misuse is crucial. This paper introduces the Graph of Attacks with Pruning (GAP) framework, an advanced approach for generating stealthy jailbreak prompts to evaluate and enhance LLM safeguards. GAP addresses limitations in existing tree-based LLM jailbreak methods by implementing an interconnected graph structure that enables knowledge sharing across attack paths. Our experimental evaluation demonstrates GAP’s superiority over existing techniques, achieving a 20.8% increase in attack success rates while reducing query costs by 62.7%. GAP consistently outperforms state-of-the-art methods for attacking both open and closed LLMs, with attack success rates of >96%. Additionally, we present specialized variants like GAP-Auto for automated seed generation and GAP-VLM for multimodal attacks. GAP-generated prompts prove highly effective in improving content moderation systems, increasing true positive detection rates by 108.5% and accuracy by 183.6% when used for fine-tuning.

The generation of novel structured data, such as graphs, with desired properties is a critical task in fields like drug discovery and material design. Graph Variational Autoencoders (GraphVAEs) offer a promising solution; however, their performance is often limited by suboptimal graph reconstruction. To address this limitation, we propose a novel framework that uses an efficient primal-dual graph matching algorithm to significantly enhance the reconstruction capabilities of GraphVAEs. By accurately aligning generated graphs with their original input graphs, our method enables the generation of valid structured data. We evaluate our approach on a benchmark molecular dataset, demonstrating its performance compared to existing graph-based generative models. Our experiments show a significant reduction in reconstruction loss and substantial improvement in training efficiency. Beyond molecular generation, the underlying metric labeling problem addressed by our graph matching algorithm has far-reaching implications for various domains, including semi-supervised learning, point cloud semantic segmentation, and the prediction of individual relations in social networks, image, and video processing. This work underscores the critical role of effective graph matching in advancing GraphVAE performance and paves the way for new possibilities in molecular generation and beyond.

Chain-of-thought (CoT) prompting is a popular in-context learning (ICL) approach for large language models (LLMs), especially when tackling complex reasoning tasks. Traditional ICL approaches construct prompts using examples that contain questions similar to the input question. However, CoT prompting, which includes crucial intermediate reasoning steps (rationales) within its examples, necessitates selecting examples based on these rationales rather than the questions themselves. Existing methods require human experts or pre-trained LLMs to describe the skill, a high-level abstraction of rationales, to guide the selection. These methods, however, are often costly and difficult to scale. Instead, this paper introduces a new approach named Latent Reasoning Skills (LaRS) that employs unsupervised learning to create a latent space representation of rationales, with a latent variable called a reasoning skill. Concurrently, LaRS learns a reasoning policy to determine the required reasoning skill for a given question. Then the ICL examples are selected by aligning the reasoning skills between past examples and the question. This approach is theoretically grounded and compute-efficient, eliminating the need for auxiliary LLM inference or manual prompt design. Empirical results demonstrate that LaRS consistently outperforms SOTA skill-based selection methods, processing example banks four times faster, reducing LLM inferences during the selection stage by half, and showing greater robustness to sub-optimal example banks.

Recent NLP literature pays little attention to the robustness of toxicity language predictors, while these systems are most likely to be used in adversarial contexts. This paper presents a novel adversarial attack, \texttt{ToxicTrap}, introducing small word-level perturbations to fool SOTA text classifiers to predict toxic text samples as benign. ToxicTrap exploits greedy based search strategies to enable fast and effective generation of toxic adversarial examples. Two novel goal function designs allow ToxicTrap to identify weaknesses in both multiclass and multilabel toxic language detectors. Our empirical results show that SOTA toxicity text classifiers are indeed vulnerable to the proposed attacks, attaining over 98\% attack success rates in multilabel cases. We also show how a vanilla adversarial training and its improved version can help increase robustness of a toxicity detector even against unseen attacks.

Modern smart card is a multi-purpose tool that is used in many areas of science and technology [5]. Microprocessor cards are used for access control, e-commerce, identification and authentication, used as personal data protected storages, as the core of e-passports and documents. In circulation there are tens of millions of devices simultaneously, and it is all the more credible by both the producers and by users [7]. All this leads to increase of attention from both developers (to protect information) and intruders. In this context, the problem of the development of new software and hardware solutions for smart cards and security analysis using their information systems (IS), including research methods, modeling, statistical estimation, as well as specialized software tools is important.

A critical component in computational treatment of an automated document labeling is the choice of an appropriate representation. Proper representation captures specific phenomena of interest in data while transforming it to a format appropriate for a classifier. For a text document, a popular choice is the bag-of-words (BoW) representation that encodes presence of unique words with non-zero weights such as TF-IDF. Extending this model to long, overlapping phrases (n-grams) results in exponential explosion in the dimensionality of the representation. In this work, we develop a model that encodes long phrases in a low-dimensional latent space with a cumulative function of individual words in each phrase. In contrast to BoW, the parameter space of the proposed model grows linearly with the length of the phrase. The proposed model requires only vector additions and multiplications with scalars to compute the latent representation of phrases, which makes it applicable to large-scale text labeling problems. Several sentiment classification and binary topic categorization problems will be used to empirically evaluate the proposed representation. The same model can also encode relative spatial distribution of elements in higher-dimensional sequences. In order to verify this claim, the proposed model will be evaluated on a large-scale image classification dataset, where images are transformed into two-dimensional sequences of quantized image descriptors.

In this paper, we introduce a novel approach for modeling n-grams in a latent space learned from supervised signals. The proposed procedure uses only unigram features to model short phrases (n-grams) in the latent space. The phrases are then combined to form document-level latent representation for a given text, where position of an n-gram in the document is used to compute corresponding combining weight. The resulting two-stage supervised embedding is then coupled with a classifier to form an end-to-end system that we apply to the large-scale sentiment classification task. The proposed model does not require feature selection to retain effective features during pre-processing, and its parameter space grows linearly with size of n-gram. We present comparative evaluations of this method using two large-scale datasets for sentiment classification in online reviews (Amazon and TripAdvisor). The proposed method outperforms standard baselines that rely on bag-of-words representation populated with n-gram features.

Spatial pyramid matching (SPM) component is part of most state-of-art image classification methods. SPM encodes spatial distribution of image features, in an un-supervised fashion, by partitioning an image into regions at multiple scales and concatenating feature vectors for these regions. In this paper we propose to replace the unsupervised SPM procedure with a supervised two-stage feature selection that requires the image partitioned at a single scale. Experimental results show the proposed method performs statistically significantly better than the SPM baseline.

In this paper, we propose an efficient embedding for modeling higher-order (n-gram) phrases that projects the n-grams to low-dimensional latent semantic space, where a classification function can be defined. We utilize a deep neural network to build a unified discriminative framework that allows for estimating the parameters of the latent space as well as the classification function with a bias for the target classification task at hand. We apply the framework to large-scale sentimental classification task. We present comparative evaluation of the proposed method on two (large) benchmark data sets for online product reviews. The proposed method achieves superior performance in comparison to the state of the art.

Concept-based representation -combined with some classifier (e. g., support vector machine) or regression analysis (e. g., linear regression)-induces a popular approach among image processing community, used to infer image labels. We propose a supervised learning procedure to obtain an embedding to a latent concept space with the pre-defined inner product. This learning procedure uses rank minimization of the sought inner product matrix, defined in the original concept space, to find an embedding to a new low dimensional space. The empirical evidence show that the proposed supervised learning method can be used in combination with another computational image embedding procedure, such as bag-of-features method, to significantly improve accuracy of label inference, while producing embedding of low complexity.

In this work we propose a novel variational method that we intend to use for estimating non-rigid texture deformation. The method is able to capture variation in gray scale images with respect to the geometry of its features. Accurate localization of features in the presence of unknown deformations is a crucial property for texture characterization. Our experimental evaluations demonstrate that accounting for geometry of features in texture images leads to significant improvements in localization of these features, when textures undergo geometrical transformations. In addition, feature descriptors using geometrical total variation energies discriminate between various regular textures with accuracy comparable to SIFT descriptors, while reduced dimensionality of TVG descriptor yields significant improvements over SIFT in terms of retrieval time.

The application of digital technologies to culture history preservation and interpretation is a rapidly growing field that has captured the imagination of many. In this work, we explore the application of image classification systems for use in the reconstruction of archaeologically excavated 18th and 19th-century ceramic fragments. In specific, we investigate the classification of thin-shell ceramics based on color and texture descriptors in order to aid in vessel reconstructions. In addition to using well known SIFT features, we formulate a new feature descriptor based on total variation geometry. The experimental results demonstrate that the proposed descriptor accurately represents texture and that, when categorizing fragments, it is best to utilize both the color and texture information available.